Safe Inputs but Unsafe Output: Benchmarking Cross-modality Safety Alignment of Large Vision-Language Model

TL;DR

This paper introduces the SIUO benchmark to evaluate cross-modality safety alignment in large vision-language models, revealing significant safety vulnerabilities and highlighting the need for improved safety measures in multi-modal AI systems.

Contribution

It presents a novel safety challenge and benchmark for cross-modality safety, addressing a gap in existing safety evaluations for multi-modal AI models.

Findings

Substantial safety vulnerabilities found in current LVLMs

Current models struggle with complex, real-world safety scenarios

Benchmark covers 9 critical safety domains

Abstract

As Artificial General Intelligence (AGI) becomes increasingly integrated into various facets of human life, ensuring the safety and ethical alignment of such systems is paramount. Previous studies primarily focus on single-modality threats, which may not suffice given the integrated and complex nature of cross-modality interactions. We introduce a novel safety alignment challenge called Safe Inputs but Unsafe Output (SIUO) to evaluate cross-modality safety alignment. Specifically, it considers cases where single modalities are safe independently but could potentially lead to unsafe or unethical outputs when combined. To empirically investigate this problem, we developed the SIUO, a cross-modality benchmark encompassing 9 critical safety domains, such as self-harm, illegal activities, and privacy violations. Our findings reveal substantial safety vulnerabilities in both closed- and open-source LVLMs, such as GPT-4V and LLaVA, underscoring the inadequacy of current models to reliably interpret and respond to complex, real-world scenarios.