Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

All Changes May Have Invariant Principles: Improving Ever-Shifting Harmful Meme Detection via Design Concept Reproduction

Ziyou Jiang, Mingyang Li, Junjie Wang, Yuekai Huang, Jie Huang, Zhiyuan Chang, Zhaoyang Li, Qing Wang

TL;DR

RepMD tackles the challenge of detecting ever-shifting harmful memes by introducing a Design Concept Graph (DCG) that encodes invariant design principles behind meme harm. It derives the DCG from historical memes via fail-reason analysis and reproduction-step derivation, then prunes it with a fast SVD-based method to enable scalable DCG-guided detection by Multimodal LLMs. On GOAT-Bench-type memes and temporally evolving Twitter memes, RepMD achieves 81.1% in-domain accuracy with modest OOD and TE gains and substantially improves human discovery efficiency (15–30 seconds per meme). The approach also provides explainable guidance to moderators and safeguards for text-to-image models, offering a principled path to robust, transferable harmful-meme detection in dynamic online environments.

Abstract

Harmful memes are ever-shifting in the Internet communities, which are difficult to analyze due to their type-shifting and temporal-evolving nature. Although these memes are shifting, we find that different memes may share invariant principles, i.e., the underlying design concept of malicious users, which can help us analyze why these memes are harmful. In this paper, we propose RepMD, an ever-shifting harmful meme detection method based on the design concept reproduction. We first refer to the attack tree to define the Design Concept Graph (DCG), which describes steps that people may take to design a harmful meme. Then, we derive the DCG from historical memes with design step reproduction and graph pruning. Finally, we use DCG to guide the Multimodal Large Language Model (MLLM) to detect harmful memes. The evaluation results show that RepMD achieves the highest accuracy with 81.1% and has slight accuracy decreases when generalized to type-shifting and temporal-evolving memes. Human evaluation shows that RepMD can improve the efficiency of human discovery on harmful memes, with 15$\sim$30 seconds per meme.

All Changes May Have Invariant Principles: Improving Ever-Shifting Harmful Meme Detection via Design Concept Reproduction

TL;DR

RepMD tackles the challenge of detecting ever-shifting harmful memes by introducing a Design Concept Graph (DCG) that encodes invariant design principles behind meme harm. It derives the DCG from historical memes via fail-reason analysis and reproduction-step derivation, then prunes it with a fast SVD-based method to enable scalable DCG-guided detection by Multimodal LLMs. On GOAT-Bench-type memes and temporally evolving Twitter memes, RepMD achieves 81.1% in-domain accuracy with modest OOD and TE gains and substantially improves human discovery efficiency (15–30 seconds per meme). The approach also provides explainable guidance to moderators and safeguards for text-to-image models, offering a principled path to robust, transferable harmful-meme detection in dynamic online environments.

Abstract

Harmful memes are ever-shifting in the Internet communities, which are difficult to analyze due to their type-shifting and temporal-evolving nature. Although these memes are shifting, we find that different memes may share invariant principles, i.e., the underlying design concept of malicious users, which can help us analyze why these memes are harmful. In this paper, we propose RepMD, an ever-shifting harmful meme detection method based on the design concept reproduction. We first refer to the attack tree to define the Design Concept Graph (DCG), which describes steps that people may take to design a harmful meme. Then, we derive the DCG from historical memes with design step reproduction and graph pruning. Finally, we use DCG to guide the Multimodal Large Language Model (MLLM) to detect harmful memes. The evaluation results show that RepMD achieves the highest accuracy with 81.1% and has slight accuracy decreases when generalized to type-shifting and temporal-evolving memes. Human evaluation shows that RepMD can improve the efficiency of human discovery on harmful memes, with 1530 seconds per meme.
Paper Structure (48 sections, 3 theorems, 10 equations, 11 figures, 7 tables, 1 algorithm)

This paper contains 48 sections, 3 theorems, 10 equations, 11 figures, 7 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Definition of Design Concept
  3. Definition of Fail Reason Tree
  4. Definition of DCG
  5. Overview of RepMD
  6. Fail Reason Tree Construction
  7. Collaborative Fail Case Extraction
  8. Collaborative Fail Reason Reflection
  9. DCG Construction
  10. Reproduction Step Node Derivation
  11. SVD-based DCG Pruning
  12. DCG-Guided Harmful Meme Detection
  13. Step-1: Relevant DCG Reproduction Step Retrieval.
  14. Step-2: Target Guidance Generation.
  15. Experimental Design
  16. ...and 33 more sections

Key Result

Theorem 1

For a random variable $X$ following a long-tail distribution, the logarithmic transform $Y = \ln(X)$ typically follows a distribution with exponentially decaying tails (e.g., normal distribution for log-normal, exponential for power-law).

Figures (11)

  • Figure 1: The motivation example of RepMD.
  • Figure 2: The overview of RepMD.
  • Figure 3: The case study of temporal evolving memes.
  • Figure 4: The results of human evaluation.
  • Figure 5: The DCG extracted from the example meme.
  • ...and 6 more figures

Theorems & Definitions (4)

  • Theorem 1: Logarithmic Normalization
  • Lemma 1: Spacing Distribution
  • Theorem 2: Maximum Gap Optimality
  • proof