Search papers, labs, and topics across Lattice.
This paper introduces MMAligner, a method designed to enhance the safety of multimodal large language models (MLLMs) by calibrating unsafe multimodal representations into a pre-existing refusal region. The authors identify that the safety disparity arises from a representation shift in unsafe multimodal inputs, which allows them to bypass intrinsic safety mechanisms. Through experiments, MMAligner achieves a 99% refusal rate on unsafe multimodal inputs with less than 2% degradation in utility, significantly improving the safety-utility trade-off compared to existing methods.
MMLMs can be made 99% safer against harmful multimodal inputs without sacrificing utility, thanks to a novel calibration approach.
Multimodal large language models (MLLMs) often refuse unsafe text prompts yet generate harmful responses to semantically equivalent multimodal inputs. Existing defenses either rely on external guardrails, which add inference overhead without repairing intrinsic flaws, or safety fine-tuning, which treats alignment as black-box optimization and may sacrifice utility or require large multimodal datasets. To identify the cause of this safety disparity, we analyze MLLM representations geometrically. We find that safety mechanisms learned from text persist across modalities: a shared safety subspace and refusal boundary remain effective, and representations inside this boundary consistently trigger refusals. However, unsafe multimodal inputs undergo a representation shift that places most of them outside the boundary, allowing them to bypass the model's intrinsic safety mechanism. This indicates that multimodal safety degradation stems from representation misalignment rather than the absence of safety capability. Based on this finding, we propose MMAligner, a safeguarding method that calibrates unsafe multimodal representations into the pre-existing refusal region. MMAligner applies a hard lower bound to ensure refusal, a soft upper bound to avoid excessive modification, and a preservation objective for benign inputs. Experiments across multiple open-source MLLMs show that MMAligner raises the average refusal rate on unsafe multimodal inputs to 99% with less than 2% utility degradation and minimal training data, substantially improving the safety-utility trade-off over existing baselines. (*Due to the notification from arXiv,"The Abstract field cannot be longer than 1,920 characters", the Abstract that appeared is shortened.)
