Relationship-Aware Safety Unlearning for Multimodal LLMs

TL;DR

This paper introduces a relationship-aware safety unlearning framework for multimodal LLMs that targets unsafe object-relation-object tuples, reducing safety failures while preserving benign relations and minimizing collateral damage.

Contribution

It proposes a novel, parameter-efficient method using LoRA to explicitly represent and unlearn unsafe relational tuples in multimodal models.

Findings

Effective suppression of unsafe object-relation-object tuples.

Robustness against paraphrase, contextual, and out-of-distribution attacks.

Preservation of safe object marginals and relations.

Abstract

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.