arXiv 2507.119682025-07-16cs.CV

观看、倾听、理解、误导：针对短视频内容适宜性评估的三模态对抗攻击

Watch, Listen, Understand, Mislead: Tri-modal Adversarial Attacks on Short Videos for Content Appropriateness Evaluation

United International University（国际大学）；BRAC University（BRAC大学）；University of British Columbia（不列颠哥伦比亚大学）；Bangladesh University of Professionals（孟加拉专业大学）；University of Alberta（阿尔伯塔大学）

Sahid Hossain Mustakim, S M Jishanul Islam, Ummay Maria Muna, Montasir Chowdhury, Mohammed Jawwadul Islam, Sadia Ahmmed, Tashfia Sikder, Syed Tasdid Azam Dhrubo, Swakkhar Shatabda

备注：: Accepted as long paper, SVU Workshop at ICCV 2025

阅读 PDF

arXiv 页面 ↗HTML 原文 ↗

AI总结：

本文提出评估MLLMs三模态安全的综合框架，包含SVMA数据集与新型三模态攻击策略ChimeraBreak，揭示模型在短视频内容审核中的漏洞与误分类偏差，为开发更安全的模型提供关键见解。

摘要：

多模态大语言模型（MLLMs）日益广泛地用于内容审核，但其在短视频场景中的鲁棒性仍未被充分探索。当前的安全评估通常依赖单模态攻击，未能解决组合攻击漏洞。本文提出了一个评估MLLMs三模态安全的综合框架。首先，我们提出了短视频多模态对抗（SVMA）数据集，包含带有由人类引导的合成对抗攻击的多样化短视频。其次，我们提出了ChimeraBreak，一种新型三模态攻击策略，同时挑战视觉、听觉和语义推理路径。在state-of-the-art MLLMs上的大量实验揭示了显著的漏洞，具有高攻击成功率（ASR）。我们的发现揭示了独特的失败模式，表明模型在将良性或违规内容误分类方面存在偏差。我们使用LLM-as-a-judge评估结果，证明了攻击推理的有效性。我们的数据集和发现为开发更鲁棒和安全的MLLMs提供了关键见解。

英文摘要：

Multimodal Large Language Models (MLLMs) are increasingly used for content moderation, yet their robustness in short-form video contexts remains underexplored. Current safety evaluations often rely on unimodal attacks, failing to address combined attack vulnerabilities. In this paper, we introduce a comprehensive framework for evaluating the tri-modal safety of MLLMs. First, we present the Short-Video Multimodal Adversarial (SVMA) dataset, comprising diverse short-form videos with human-guided synthetic adversarial attacks. Second, we propose ChimeraBreak, a novel tri-modal attack strategy that simultaneously challenges visual, auditory, and semantic reasoning pathways. Extensive experiments on state-of-the-art MLLMs reveal significant vulnerabilities with high Attack Success Rates (ASR). Our findings uncover distinct failure modes, showing model biases toward misclassifying benign or policy-violating content. We assess results using LLM-as-a-judge, demonstrating attack reasoning efficacy. Our dataset and findings provide crucial insights for developing more robust and safe MLLMs.

机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。