Rainbow Noise: Stress-Testing Multimodal Harmful-Meme Detectors on LGBTQ Content

2025-07-24

AI Summary

Harmful memes targeting LGBTQ+ communities frequently evade existing detection models via textual obfuscation and image perturbations, exposing critical robustness deficiencies in current multimodal approaches. To address this, we introduce the first LGBTQ+-focused multimodal harmful meme robustness benchmark, incorporating four categories of textual attacks and three types of image perturbations. We further propose a lightweight Text Denoising Adapter (TDA) that significantly enhances model resilience against adversarial text variants. Systematic adversarial evaluations are conducted on MemeCLIP and MemeBLIP2. Ablation studies identify architecture design and data composition as key determinants of robustness. Results show that MemeCLIP exhibits relatively stable robustness, whereas MemeBLIP2 integrated with TDA achieves superior performance—demonstrating that lightweight, modular enhancements can effectively strengthen the adversarial resilience of multimodal safety systems.

Technology Category

• Natural Language Processing: Safety and Robustness
• Machine Learning: Adversarial Learning & Robustness
• Computer Vision: Adversarial Attacks & Robustness

Application Category

• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Responsible Web: Mitigating online harms, including harassment, spam, and fake reviews

Abstract

Hateful memes aimed at LGBTQ,+ communities often evade detection by tweaking either the caption, the image, or both. We build the first robustness benchmark for this setting, pairing four realistic caption attacks with three canonical image corruptions and testing all combinations on the PrideMM dataset. Two state-of-the-art detectors, MemeCLIP and MemeBLIP2, serve as case studies, and we introduce a lightweight extbf{Text Denoising Adapter (TDA)} to enhance the latter's resilience. Across the grid, MemeCLIP degrades more gently, while MemeBLIP2 is particularly sensitive to the caption edits that disrupt its language processing. However, the addition of the TDA not only remedies this weakness but makes MemeBLIP2 the most robust model overall. Ablations reveal that all systems lean heavily on text, but architectural choices and pre-training data significantly impact robustness. Our benchmark exposes where current multimodal safety models crack and demonstrates that targeted, lightweight modules like the TDA offer a powerful path towards stronger defences.

Problem

• Detecting hateful memes targeting LGBTQ+ communities effectively
• Testing robustness of multimodal detectors against text and image attacks
• Improving resilience of detectors with lightweight text denoising adapters

Innovation

• First robustness benchmark for hateful meme detection
• Lightweight Text Denoising Adapter enhances resilience
• Combines caption attacks with image corruptions testing