Automated Safety Benchmarking: A Multi-agent Pipeline for LVLMs

TL;DR

This paper introduces VLSafetyBencher, an automated multi-agent system that rapidly constructs high-quality safety benchmarks for LVLMs, addressing limitations of manual benchmarks and enhancing model safety evaluation.

Contribution

The paper presents the first automated system for LVLM safety benchmarking, significantly reducing construction time and improving discriminative power of safety assessments.

Findings

Constructs high-quality safety benchmarks within one week

Achieves a 70% safety rate disparity among models

Demonstrates effectiveness of multi-agent approach in safety evaluation

Abstract

Large vision-language models (LVLMs) exhibit remarkable capabilities in cross-modal tasks but face significant safety challenges, which undermine their reliability in real-world applications. Efforts have been made to build LVLM safety evaluation benchmarks to uncover their vulnerability. However, existing benchmarks are hindered by their labor-intensive construction process, static complexity, and limited discriminative power. Thus, they may fail to keep pace with rapidly evolving models and emerging risks. To address these limitations, we propose VLSafetyBencher, the first automated system for LVLM safety benchmarking. VLSafetyBencher introduces four collaborative agents: Data Preprocessing, Generation, Augmentation, and Selection agents to construct and select high-quality samples. Experiments validates that VLSafetyBencher can construct high-quality safety benchmarks within one week at a minimal cost. The generated benchmark effectively distinguish safety, with a safety rate disparity of 70% between the most and least safe models.