Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Ran Tong, Songtao Wei, Jiaqi Liu, Lanruo Wang

TL;DR

This work addresses the vulnerability of multimodal hate-meme detectors to realistic text and image perturbations targeting LGBTQ content. It constructs the PrideMM robustness benchmark, evaluates MemeCLIP and MemeBLIP2 (including a new MemeBLIP2+TDA variant), and demonstrates that text is the dominant information channel while image perturbations have weaker impact. A lightweight Text Denoising Adapter (TDA) is proposed and shown to significantly improve MemeBLIP2’s robustness under combined perturbations, even making it the most robust model among those tested. The study also introduces PrideMM-Aug and provides a rigorous framework of robustness metrics and ablations to guide future defenses in multimodal safety systems.

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 \textbf{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.

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

TL;DR

This work addresses the vulnerability of multimodal hate-meme detectors to realistic text and image perturbations targeting LGBTQ content. It constructs the PrideMM robustness benchmark, evaluates MemeCLIP and MemeBLIP2 (including a new MemeBLIP2+TDA variant), and demonstrates that text is the dominant information channel while image perturbations have weaker impact. A lightweight Text Denoising Adapter (TDA) is proposed and shown to significantly improve MemeBLIP2’s robustness under combined perturbations, even making it the most robust model among those tested. The study also introduces PrideMM-Aug and provides a rigorous framework of robustness metrics and ablations to guide future defenses in multimodal safety systems.

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 \textbf{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.

Paper Structure

This paper contains 23 sections, 14 equations, 1 figure, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Image Noise Generation
  5. Textual noise generation
  6. Enhancing Textual Robustness with an Adaptive Module
  7. Experiments
  8. Experimental Setup
  9. Evaluation Metrics
  10. Results for the image part
  11. Results for the text part
  12. Which Input Channel Drives Performance?
  13. Clean conditions
  14. Image perturbations
  15. Combined text–image perturbations
  16. ...and 8 more sections

Figures (1)

  • Figure 1: Original meme (top left) and its three noisy variants: Universal adversarial perturbations (top right), Common corruptions and perturbations (bottom left), and AugMix compositional augmentation (bottom right).