When Harmful Content Gets Camouflaged: Unveiling Perception Failure of LVLMs with CamHarmTI

2025/11/29 by Yanhui Li, Qi Zhou, Li, Yanhui +9
Computer Science · Social Sciences · #Artificial Intelligence (cs.AI) #FOS: Computer and information sciences #Generative Adversarial Networks and Image Synthesis #Misinformation and Its Impacts #Multimedia (cs.MM) #Multimodal Machine Learning Applications

paper · pdf · doi:10.48550/arxiv.2512.03087

openalex publication_date 2025/11/29 · openalex created_date 2025/12/05 · openalex updated_date 2026/07/28

Abstract

Large vision-language models (LVLMs) are increasingly used for tasks where detecting multimodal harmful content is crucial, such as online content moderation. However, real-world harmful content is often camouflaged, relying on nuanced text-image interplay, such as memes or images with embedded malicious text, to evade detection. This raises a key question: can LVLMs perceive such camouflaged harmful content as sensitively as humans do? In this paper, we introduce CamHarmTI, a benchmark for evaluating LVLM ability to perceive and interpret camouflaged harmful content within text-image compositions. CamHarmTI consists of over 4,500 samples across three types of image-text posts. Experiments on 100 human users and 12 mainstream LVLMs reveal a clear perceptual gap: humans easily recognize such content (e.g., over 95.75% accuracy), whereas current LVLMs often fail (e.g., ChatGPT-4o achieves only 2.10% accuracy). Moreover, fine-tuning experiments demonstrate that \bench serves as an effective resource for improving model perception, increasing accuracy by 55.94% for Qwen2.5VL-7B. Attention analysis and layer-wise probing further reveal that fine-tuning enhances sensitivity primarily in the early layers of the vision encoder, promoting a more integrated scene understanding. These findings highlight the inherent perceptual limitations in LVLMs and offer insight into more human-aligned visual reasoning systems.

Citations

• On the Reliability of Vision-Language Models Under Adversarial Frequency-Domain Perturbations
• Position: Reasoning After Perception Means Reasoning Without Vision
• Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities
• Emerging Properties in Unified Multimodal Pretraining
• Cross-Platform Violence Detection on Social Media: A Dataset and Analysis
• Unveiling the Lack of LVLM Robustness to Fundamental Visual Variations: Why and Path Forward
• Analysing the Robustness of Vision-Language-Models to Common Corruptions
• Kimi-VL Technical Report
• Gemma 3 Technical Report
• A Framework for a Capability-driven Evaluation of Scenario Understanding for Multimodal Large Language Models in Autonomous Driving
• PTDiffusion: Free Lunch for Generating Optical Illusion Hidden Pictures with Phase-Transferred Diffusion Model
• Qwen2.5-VL Technical Report
• IllusionCAPTCHA: A CAPTCHA based on Visual Illusion
• Janus-Pro: Unified Multimodal Understanding and Generation with Data and Model Scaling
• Towards Safer Social Media Platforms: Scalable and Performant Few-Shot Harmful Content Moderation Using Large Language Models
• Early evidence of how LLMs outperform traditional systems on OCR/HTR tasks for historical records
• Religious Bias Landscape in Language and Text-to-Image Models: Analysis, Detection, and Debiasing Strategies
• ICM-Assistant: Instruction-tuning Multimodal Large Language Models for Rule-based Explainable Image Content Moderation
• Comprehensive Evaluation of Multimodal AI Models in Medical Imaging Diagnosis: From Data Augmentation to Preference-Based Comparison
• Sensitive Content Classification in Social Media: A Holistic Resource and Evaluation
• MME-Survey: A Comprehensive Survey on Evaluation of Multimodal LLMs
• Llama Guard 3 Vision: Safeguarding Human-AI Image Understanding Conversations
• Jailbreak Attacks and Defenses against Multimodal Generative Models: A Survey
• A Comprehensive Survey and Guide to Multimodal Large Language Models in Vision-Language Tasks
• GPT-4o System Card
• TrojVLM: Backdoor Attack Against Vision Language Models
• Content Moderation by LLM: From Accuracy to Legitimacy
• MM-Vet v2: A Challenging Benchmark to Evaluate Large Multimodal Models for Integrated Capabilities
• Revisiting Backdoor Attacks against Large Vision-Language Models from Domain Shift
• SEED-Bench-2-Plus: Benchmarking Multimodal Large Language Models with Text-Rich Visual Comprehension
• LLMGuard: Guarding Against Unsafe LLM Behavior
• Spy-Watermark: Robust Invisible Watermarking for Backdoor Attack
• Improved Baselines with Visual Instruction Tuning
• Adversarial Illusions in Multi-Modal Embeddings
• MMBench: Is Your Multi-modal Model an All-around Player?
• Vision-Language Models for Vision Tasks: A Survey
• Vision-Language Models for Vision Tasks: A Survey
• Multilingual Content Moderation: A Case Study on Reddit
• MMBench: Benchmarking End-to-End Multi-modal DNNs and Understanding Their Hardware-Software Implications
• Hierarchical Text-Conditional Image Generation with CLIP Latents
• Denoising Diffusion Probabilistic Models
• The Hateful Memes Challenge: Detecting Hate Speech in Multimodal Memes
• Algorithmic content moderation: Technical and political challenges in the automation of platform governance
• Recurrent Neural Networks (RNNs): A gentle Introduction and Overview
• Can We Talk Models Into Seeing the World Differently?
• Recent Advances in Hate Speech Moderation: Multimodality and the Role of Large Models

Related

• B-AVIBench: Towards Evaluating the Robustness of Large Vision-Language Model on Black-box Adversarial Visual-Instructions
• REVEAL: Multi-turn Evaluation of Image-Input Harms for Vision LLM
• Playing Devil's Advocate: Unmasking Toxicity and Vulnerabilities in Large Vision-Language Models
• Are Vision-Language Models Safe in the Wild? A Meme-Based Benchmark Study
• SMSP: A Plug-and-Play Strategy of Multi-Scale Perception for MLLMs to Perceive Visual Illusions