Cross-Modal Safety Mechanism Transfer in Large Vision-Language Models

2024-10-16

arXiv.org

AI Summary

Existing large vision-language models (LVLMs) struggle to transfer mature textual safety mechanisms from large language models (LLMs) to the visual modality, resulting in ineffective detection and blocking of toxic images. This work is the first to identify that LVLM safety responses concentrate in specific Transformer layer hidden states. Leveraging this insight, we propose Text-Guided hidden-state Alignment (TGA): a zero-shot cross-modal safety transfer method that uses safe text queries to guide the projection of visual features into the safety-critical hidden-state subspace—without fine-tuning the vision encoder. TGA preserves performance on diverse downstream tasks—including visual question answering and image captioning—while significantly improving toxic image detection accuracy. Our approach thus achieves the dual objective of “Safe and Good”: robust safety enforcement without compromising multimodal understanding capabilities.

Technology Category

• Computer Vision: Large Vision Models
• Natural Language Processing: Safety and Robustness
• Machine Learning: Large Multimodal Models (LMMs)

Application Category

• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• Security and Privacy: Large-scale security measurements
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Vision-language alignment in Large Vision-Language Models (LVLMs) successfully enables LLMs to understand visual input. However, we find that existing vision-language alignment methods fail to transfer the existing safety mechanism for text in LLMs to vision, which leads to vulnerabilities in toxic image. To explore the cause of this problem, we give the insightful explanation of where and how the safety mechanism of LVLMs operates and conduct comparative analysis between text and vision. We find that the hidden states at the specific transformer layers play a crucial role in the successful activation of safety mechanism, while the vision-language alignment at hidden states level in current methods is insufficient. This results in a semantic shift for input images compared to text in hidden states, therefore misleads the safety mechanism. To address this, we propose a novel Text-Guided vision-language Alignment method (TGA) for LVLMs. TGA retrieves the texts related to input vision and uses them to guide the projection of vision into the hidden states space in LLMs. Experiments show that TGA not only successfully transfers the safety mechanism for text in basic LLMs to vision in vision-language alignment for LVLMs without any safety fine-tuning on the visual modality but also maintains the general performance on various vision tasks (Safe and Good).

Problem

• Transfer safety mechanisms from text to vision in LVLMs
• Address vulnerabilities in toxic image understanding
• Improve vision-language alignment at hidden states level

Innovation

• Text-Guided vision-language Alignment method (TGA)
• Retrieves texts to guide vision projection
• Transfers text safety mechanisms to vision